SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers

We present SegFormer, a simple, efficient yet powerful semantic segmentation which unifies Transformers with lightweight multilayer perception (MLP) decoders. SegFormer has two appealing features: 1) SegFormer comprises a hierarchically structured Transformer encoder which outputs multiscale. It does not need positional encoding, thereby avoiding the of positional codes which leads to decreased performance when the resolution differs from training. 2) SegFormer avoids complex decoders. proposed MLP decoder aggregates information from different layers, and thus both local attention and global attention to render powerful. We show that this simple and lightweight design is the key to segmentation on Transformers. We scale our approach up to obtain a of models from SegFormer-B0 to SegFormer-B5, reaching significantly performance and efficiency than previous counterparts. For example, -B4 achieves 50. 3% mIoU on ADE20K with 64M parameters, being 5x and 2. 2% better than the previous best method. Our best model, -B5, achieves 84. 0% mIoU on Cityscapes validation set and shows zero-shot robustness on Cityscapes-C. Code will be released at: . com/NVlabs/SegFormer.